GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification

Implementation of KDD'23 paper GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification.

Requirements

This repository has been tested with the following packages:

• Python == 3.8.13
• PyTorch == 1.11.0
• PyTorch Geometric == 2.0.4

Please follow official instructions to install Pytorch and Pytorch Geometric.

Important Hyper-parameters

• --dataset: name of the dataset. Could be one of ['Cora', 'CiteSeer', 'PubMed', 'Amazon-Photo', 'Amazon-Computers', 'Coauthor-CS'].
• --data_path: the path to the dataset. The dataset will be downloaded to this path automatically when the code is executed for the first time.
• --imb_ratio: imbalance ratio.
• --net: GNN backbone. Could be one of ['GCN, GAT, SAGE'].
• --gdc: way to get the weighted graph. Could be one of ['ppr', 'hk', 'none'].

Please refer to args.py for the full hyper-parameters.

How to Run

Pass the above hyper-parameters to main.py. For example:

python main.py --dataset Cora --data_path dataset/ --imb_ratio 100 --net GCN --gdc ppr
python main.py --dataset CiteSeer --data_path dataset/ --imb_ratio 100 --net GCN --gdc ppr --max
python main.py --dataset PubMed --data_path dataset/ --imb_ratio 100 --net GCN --gdc none --max --no_mask
python main.py --dataset Amazon-Photo --data_path dataset/ --imb_ratio 20 --net SAGE --gdc ppr --max
python main.py --dataset Amazon-Computers --data_path dataset/ --imb_ratio 20 --net SAGE --gdc ppr
python main.py --dataset Coauthor-CS --data_path dataset/ --imb_ratio 20 --net SAGE --gdc ppr --max

License

MIT License

Contact

Feel free to email (liwzh63 [AT] mail2.sysu.edu.cn) for any questions about this work.

Acknowledgements

The code is implemented based on GraphENS and ReNode.

Citation

If you find this work is helpful to your research, please consider citing our paper:

@article{li2023graphsha,
  title={GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification},
  author={Wen-Zhi Li and Chang-Dong Wang and Hui Xiong and Jian-Huang Lai},
  journal={arXiv preprint arXiv:2306.09612},
  year={2023}
}